LoRA Stack

LoRA Stack (class AtomLoRAStackEntry) is the building block of the Atom_AI stack system: a single chainable slot that holds one LoRA and its two strengths. The whole idea is legibility - instead of one node packed with ten dropdowns, you get ten little nodes, each showing its own LoRA name, and you wire them into a chain. The LoRA Stack Apply node at the end reads the chain and applies everything.

If that sounds like the opposite design from rgthree's Power Lora Loader (one node, everything inside), it is. Different tastes, same result. This one's for people who want the graph itself to be the documentation.

How it works

Each entry is pure data plumbing - it never touches a model. When you pick a LoRA and set its strengths, the node appends a small record (lora_name, strength_model, strength_clip) to a list. The list starts at the first entry, and each entry passes the growing list along through its optional lora_stack input, so the final entry's output is the complete stack.

Two details worth knowing. First, the enabled toggle and the "None" option: a disabled entry (or one left on "None") is simply skipped - it doesn't append anything, but it still passes the list through, so you can keep a slot in place and switch it off without rewiring. Second, order: the chain's order is the fusion order at apply time. First entry gets applied first. The community rule of thumb is to put the big structural LoRAs (style, base-character) near the start and subtle tweaks after - reordering is cheap to test, so don't overthink it up front.

Inputs and outputs

• lora_name - a dropdown of everything in your models/loras folder, plus "None".
• strength_model - slider from −10 to 10. Below 1.0 weakens the LoRA; above 1.0 pushes it hard; a negative value inverts it, which is occasionally useful for fighting an effect. Default 1.0 is often too strong for style LoRAs - 0.5–0.8 is the common sweet spot.
• strength_clip - same range, controls how much the text-encoder side of the LoRA changes your prompt's interpretation. For many LoRAs you can leave it at 1.0.
• enabled - the on/off toggle.
• lora_stack (optional) - the chain input. Feed it the previous entry's output; leave it empty on the first entry.

The single output is lora_stack, which you chain into the next entry (or into LoRA Stack Apply to fuse).

Installing it

This is one of six nodes in the Atom_AI pack. Install the pack once via ComfyUI Manager (search Atom_AI) or:

cd ComfyUI/custom_nodes
git clone https://github.com/AtomAI-GR/Atom_AI

Then restart ComfyUI. No extra Python dependencies and no model downloads - you only need your LoRA files where they already live in models/loras. Look for it under Atom_AI ▸ LoRA.

Common issues

• Stack "doesn't work." Usually it's not the stack - it's a LoRA that doesn't match your base model, or a trigger word that's missing from the prompt. Those two account for most "my LoRA does nothing" reports, and no loader node can fix either. Test the LoRA solo at strength 1.0 with a fixed seed before you suspect the chain.
• Entry shows but isn't applied. Check the enabled toggle - a disabled slot silently drops out. It's the node's feature, but it's also the thing that trips people who saved a workflow with half their LoRAs off.
• Order surprises. Because the chain order is fusion order, adding an entry at the wrong end changes the result. When output goes muddy, move the suspect LoRA earlier or later and re-run.

The trap to avoid is building a 12-LoRA stack and blaming the tooling when it looks overcooked. Stacking is compounding - every LoRA raises the chance the batch fights itself. If the result is mush, cut the stack to two and add back one at a time. That's true no matter whose stack nodes you use.